drivers/gpu/drm/i915/i915_vma.c

   1 /*
   2  * Copyright © 2016 Intel Corporation
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  *
  23  */
  24
  25 #include <linux/sched/mm.h>
  26 #include <drm/drm_gem.h>
  27
  28 #include "display/intel_frontbuffer.h"
  29
  30 #include "gt/intel_engine.h"
  31 #include "gt/intel_gt.h"
  32
  33 #include "i915_drv.h"
  34 #include "i915_globals.h"
  35 #include "i915_trace.h"
  36 #include "i915_vma.h"
  37
  38 static struct i915_global_vma {
  39         struct i915_global base;
  40         struct kmem_cache *slab_vmas;
  41 } global;
  42
  43 struct i915_vma *i915_vma_alloc(void)
  44 {
  45         return kmem_cache_zalloc(global.slab_vmas, GFP_KERNEL);
  46 }
  47
  48 void i915_vma_free(struct i915_vma *vma)
  49 {
  50         return kmem_cache_free(global.slab_vmas, vma);
  51 }
  52
  53 #if IS_ENABLED(CONFIG_DRM_I915_ERRLOG_GEM) && IS_ENABLED(CONFIG_DRM_DEBUG_MM)
  54
  55 #include <linux/stackdepot.h>
  56
  57 static void vma_print_allocator(struct i915_vma *vma, const char *reason)
  58 {
  59         unsigned long *entries;
  60         unsigned int nr_entries;
  61         char buf[512];
  62
  63         if (!vma->node.stack) {
  64                 DRM_DEBUG_DRIVER("vma.node [%08llx + %08llx] %s: unknown owner\n",
  65                                  vma->node.start, vma->node.size, reason);
  66                 return;
  67         }
  68
  69         nr_entries = stack_depot_fetch(vma->node.stack, &entries);
  70         stack_trace_snprint(buf, sizeof(buf), entries, nr_entries, 0);
  71         DRM_DEBUG_DRIVER("vma.node [%08llx + %08llx] %s: inserted at %s\n",
  72                          vma->node.start, vma->node.size, reason, buf);
  73 }
  74
  75 #else
  76
  77 static void vma_print_allocator(struct i915_vma *vma, const char *reason)
  78 {
  79 }
  80
  81 #endif
  82
  83 static inline struct i915_vma *active_to_vma(struct i915_active *ref)
  84 {
  85         return container_of(ref, typeof(struct i915_vma), active);
  86 }
  87
  88 static int __i915_vma_active(struct i915_active *ref)
  89 {
  90         return i915_vma_tryget(active_to_vma(ref)) ? 0 : -ENOENT;
  91 }
  92
  93 static void __i915_vma_retire(struct i915_active *ref)
  94 {
  95         i915_vma_put(active_to_vma(ref));
  96 }
  97
  98 static struct i915_vma *
  99 vma_create(struct drm_i915_gem_object *obj,
 100            struct i915_address_space *vm,
 101            const struct i915_ggtt_view *view)
 102 {
 103         struct i915_vma *vma;
 104         struct rb_node *rb, **p;
 105
 106         /* The aliasing_ppgtt should never be used directly! */
 107         GEM_BUG_ON(vm == &vm->i915->ggtt.alias->vm);
 108
 109         vma = i915_vma_alloc();
 110         if (vma == NULL)
 111                 return ERR_PTR(-ENOMEM);
 112
 113         vma->vm = vm;
 114         vma->ops = &vm->vma_ops;
 115         vma->obj = obj;
 116         vma->resv = obj->base.resv;
 117         vma->size = obj->base.size;
 118         vma->display_alignment = I915_GTT_MIN_ALIGNMENT;
 119
 120         i915_active_init(vm->i915, &vma->active,
 121                          __i915_vma_active, __i915_vma_retire);
 122
 123         /* Declare ourselves safe for use inside shrinkers */
 124         if (IS_ENABLED(CONFIG_LOCKDEP)) {
 125                 fs_reclaim_acquire(GFP_KERNEL);
 126                 might_lock(&vma->active.mutex);
 127                 fs_reclaim_release(GFP_KERNEL);
 128         }
 129
 130         INIT_LIST_HEAD(&vma->closed_link);
 131
 132         if (view && view->type != I915_GGTT_VIEW_NORMAL) {
 133                 vma->ggtt_view = *view;
 134                 if (view->type == I915_GGTT_VIEW_PARTIAL) {
 135                         GEM_BUG_ON(range_overflows_t(u64,
 136                                                      view->partial.offset,
 137                                                      view->partial.size,
 138                                                      obj->base.size >> PAGE_SHIFT));
 139                         vma->size = view->partial.size;
 140                         vma->size <<= PAGE_SHIFT;
 141                         GEM_BUG_ON(vma->size > obj->base.size);
 142                 } else if (view->type == I915_GGTT_VIEW_ROTATED) {
 143                         vma->size = intel_rotation_info_size(&view->rotated);
 144                         vma->size <<= PAGE_SHIFT;
 145                 } else if (view->type == I915_GGTT_VIEW_REMAPPED) {
 146                         vma->size = intel_remapped_info_size(&view->remapped);
 147                         vma->size <<= PAGE_SHIFT;
 148                 }
 149         }
 150
 151         if (unlikely(vma->size > vm->total))
 152                 goto err_vma;
 153
 154         GEM_BUG_ON(!IS_ALIGNED(vma->size, I915_GTT_PAGE_SIZE));
 155
 156         if (i915_is_ggtt(vm)) {
 157                 if (unlikely(overflows_type(vma->size, u32)))
 158                         goto err_vma;
 159
 160                 vma->fence_size = i915_gem_fence_size(vm->i915, vma->size,
 161                                                       i915_gem_object_get_tiling(obj),
 162                                                       i915_gem_object_get_stride(obj));
 163                 if (unlikely(vma->fence_size < vma->size || /* overflow */
 164                              vma->fence_size > vm->total))
 165                         goto err_vma;
 166
 167                 GEM_BUG_ON(!IS_ALIGNED(vma->fence_size, I915_GTT_MIN_ALIGNMENT));
 168
 169                 vma->fence_alignment = i915_gem_fence_alignment(vm->i915, vma->size,
 170                                                                 i915_gem_object_get_tiling(obj),
 171                                                                 i915_gem_object_get_stride(obj));
 172                 GEM_BUG_ON(!is_power_of_2(vma->fence_alignment));
 173
 174                 vma->flags |= I915_VMA_GGTT;
 175         }
 176
 177         spin_lock(&obj->vma.lock);
 178
 179         rb = NULL;
 180         p = &obj->vma.tree.rb_node;
 181         while (*p) {
 182                 struct i915_vma *pos;
 183                 long cmp;
 184
 185                 rb = *p;
 186                 pos = rb_entry(rb, struct i915_vma, obj_node);
 187
 188                 /*
 189                  * If the view already exists in the tree, another thread
 190                  * already created a matching vma, so return the older instance
 191                  * and dispose of ours.
 192                  */
 193                 cmp = i915_vma_compare(pos, vm, view);
 194                 if (cmp == 0) {
 195                         spin_unlock(&obj->vma.lock);
 196                         i915_vma_free(vma);
 197                         return pos;
 198                 }
 199
 200                 if (cmp < 0)
 201                         p = &rb->rb_right;
 202                 else
 203                         p = &rb->rb_left;
 204         }
 205         rb_link_node(&vma->obj_node, rb, p);
 206         rb_insert_color(&vma->obj_node, &obj->vma.tree);
 207
 208         if (i915_vma_is_ggtt(vma))
 209                 /*
 210                  * We put the GGTT vma at the start of the vma-list, followed
 211                  * by the ppGGTT vma. This allows us to break early when
 212                  * iterating over only the GGTT vma for an object, see
 213                  * for_each_ggtt_vma()
 214                  */
 215                 list_add(&vma->obj_link, &obj->vma.list);
 216         else
 217                 list_add_tail(&vma->obj_link, &obj->vma.list);
 218
 219         spin_unlock(&obj->vma.lock);
 220
 221         mutex_lock(&vm->mutex);
 222         list_add(&vma->vm_link, &vm->unbound_list);
 223         mutex_unlock(&vm->mutex);
 224
 225         return vma;
 226
 227 err_vma:
 228         i915_vma_free(vma);
 229         return ERR_PTR(-E2BIG);
 230 }
 231
 232 static struct i915_vma *
 233 vma_lookup(struct drm_i915_gem_object *obj,
 234            struct i915_address_space *vm,
 235            const struct i915_ggtt_view *view)
 236 {
 237         struct rb_node *rb;
 238
 239         rb = obj->vma.tree.rb_node;
 240         while (rb) {
 241                 struct i915_vma *vma = rb_entry(rb, struct i915_vma, obj_node);
 242                 long cmp;
 243
 244                 cmp = i915_vma_compare(vma, vm, view);
 245                 if (cmp == 0)
 246                         return vma;
 247
 248                 if (cmp < 0)
 249                         rb = rb->rb_right;
 250                 else
 251                         rb = rb->rb_left;
 252         }
 253
 254         return NULL;
 255 }
 256
 257 /**
 258  * i915_vma_instance - return the singleton instance of the VMA
 259  * @obj: parent &struct drm_i915_gem_object to be mapped
 260  * @vm: address space in which the mapping is located
 261  * @view: additional mapping requirements
 262  *
 263  * i915_vma_instance() looks up an existing VMA of the @obj in the @vm with
 264  * the same @view characteristics. If a match is not found, one is created.
 265  * Once created, the VMA is kept until either the object is freed, or the
 266  * address space is closed.
 267  *
 268  * Must be called with struct_mutex held.
 269  *
 270  * Returns the vma, or an error pointer.
 271  */
 272 struct i915_vma *
 273 i915_vma_instance(struct drm_i915_gem_object *obj,
 274                   struct i915_address_space *vm,
 275                   const struct i915_ggtt_view *view)
 276 {
 277         struct i915_vma *vma;
 278
 279         GEM_BUG_ON(view && !i915_is_ggtt(vm));
 280         GEM_BUG_ON(vm->closed);
 281
 282         spin_lock(&obj->vma.lock);
 283         vma = vma_lookup(obj, vm, view);
 284         spin_unlock(&obj->vma.lock);
 285
 286         /* vma_create() will resolve the race if another creates the vma */
 287         if (unlikely(!vma))
 288                 vma = vma_create(obj, vm, view);
 289
 290         GEM_BUG_ON(!IS_ERR(vma) && i915_vma_compare(vma, vm, view));
 291         return vma;
 292 }
 293
 294 /**
 295  * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space.
 296  * @vma: VMA to map
 297  * @cache_level: mapping cache level
 298  * @flags: flags like global or local mapping
 299  *
 300  * DMA addresses are taken from the scatter-gather table of this object (or of
 301  * this VMA in case of non-default GGTT views) and PTE entries set up.
 302  * Note that DMA addresses are also the only part of the SG table we care about.
 303  */
 304 int i915_vma_bind(struct i915_vma *vma, enum i915_cache_level cache_level,
 305                   u32 flags)
 306 {
 307         u32 bind_flags;
 308         u32 vma_flags;
 309         int ret;
 310
 311         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 312         GEM_BUG_ON(vma->size > vma->node.size);
 313
 314         if (GEM_DEBUG_WARN_ON(range_overflows(vma->node.start,
 315                                               vma->node.size,
 316                                               vma->vm->total)))
 317                 return -ENODEV;
 318
 319         if (GEM_DEBUG_WARN_ON(!flags))
 320                 return -EINVAL;
 321
 322         bind_flags = 0;
 323         if (flags & PIN_GLOBAL)
 324                 bind_flags |= I915_VMA_GLOBAL_BIND;
 325         if (flags & PIN_USER)
 326                 bind_flags |= I915_VMA_LOCAL_BIND;
 327
 328         vma_flags = vma->flags & (I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND);
 329         if (flags & PIN_UPDATE)
 330                 bind_flags |= vma_flags;
 331         else
 332                 bind_flags &= ~vma_flags;
 333         if (bind_flags == 0)
 334                 return 0;
 335
 336         GEM_BUG_ON(!vma->pages);
 337
 338         trace_i915_vma_bind(vma, bind_flags);
 339         ret = vma->ops->bind_vma(vma, cache_level, bind_flags);
 340         if (ret)
 341                 return ret;
 342
 343         vma->flags |= bind_flags;
 344         return 0;
 345 }
 346
 347 void __iomem *i915_vma_pin_iomap(struct i915_vma *vma)
 348 {
 349         void __iomem *ptr;
 350         int err;
 351
 352         /* Access through the GTT requires the device to be awake. */
 353         assert_rpm_wakelock_held(&vma->vm->i915->runtime_pm);
 354
 355         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 356         if (WARN_ON(!i915_vma_is_map_and_fenceable(vma))) {
 357                 err = -ENODEV;
 358                 goto err;
 359         }
 360
 361         GEM_BUG_ON(!i915_vma_is_ggtt(vma));
 362         GEM_BUG_ON((vma->flags & I915_VMA_GLOBAL_BIND) == 0);
 363
 364         ptr = vma->iomap;
 365         if (ptr == NULL) {
 366                 ptr = io_mapping_map_wc(&i915_vm_to_ggtt(vma->vm)->iomap,
 367                                         vma->node.start,
 368                                         vma->node.size);
 369                 if (ptr == NULL) {
 370                         err = -ENOMEM;
 371                         goto err;
 372                 }
 373
 374                 vma->iomap = ptr;
 375         }
 376
 377         __i915_vma_pin(vma);
 378
 379         err = i915_vma_pin_fence(vma);
 380         if (err)
 381                 goto err_unpin;
 382
 383         i915_vma_set_ggtt_write(vma);
 384         return ptr;
 385
 386 err_unpin:
 387         __i915_vma_unpin(vma);
 388 err:
 389         return IO_ERR_PTR(err);
 390 }
 391
 392 void i915_vma_flush_writes(struct i915_vma *vma)
 393 {
 394         if (!i915_vma_has_ggtt_write(vma))
 395                 return;
 396
 397         intel_gt_flush_ggtt_writes(vma->vm->gt);
 398
 399         i915_vma_unset_ggtt_write(vma);
 400 }
 401
 402 void i915_vma_unpin_iomap(struct i915_vma *vma)
 403 {
 404         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 405
 406         GEM_BUG_ON(vma->iomap == NULL);
 407
 408         i915_vma_flush_writes(vma);
 409
 410         i915_vma_unpin_fence(vma);
 411         i915_vma_unpin(vma);
 412 }
 413
 414 void i915_vma_unpin_and_release(struct i915_vma **p_vma, unsigned int flags)
 415 {
 416         struct i915_vma *vma;
 417         struct drm_i915_gem_object *obj;
 418
 419         vma = fetch_and_zero(p_vma);
 420         if (!vma)
 421                 return;
 422
 423         obj = vma->obj;
 424         GEM_BUG_ON(!obj);
 425
 426         i915_vma_unpin(vma);
 427         i915_vma_close(vma);
 428
 429         if (flags & I915_VMA_RELEASE_MAP)
 430                 i915_gem_object_unpin_map(obj);
 431
 432         i915_gem_object_put(obj);
 433 }
 434
 435 bool i915_vma_misplaced(const struct i915_vma *vma,
 436                         u64 size, u64 alignment, u64 flags)
 437 {
 438         if (!drm_mm_node_allocated(&vma->node))
 439                 return false;
 440
 441         if (vma->node.size < size)
 442                 return true;
 443
 444         GEM_BUG_ON(alignment && !is_power_of_2(alignment));
 445         if (alignment && !IS_ALIGNED(vma->node.start, alignment))
 446                 return true;
 447
 448         if (flags & PIN_MAPPABLE && !i915_vma_is_map_and_fenceable(vma))
 449                 return true;
 450
 451         if (flags & PIN_OFFSET_BIAS &&
 452             vma->node.start < (flags & PIN_OFFSET_MASK))
 453                 return true;
 454
 455         if (flags & PIN_OFFSET_FIXED &&
 456             vma->node.start != (flags & PIN_OFFSET_MASK))
 457                 return true;
 458
 459         return false;
 460 }
 461
 462 void __i915_vma_set_map_and_fenceable(struct i915_vma *vma)
 463 {
 464         bool mappable, fenceable;
 465
 466         GEM_BUG_ON(!i915_vma_is_ggtt(vma));
 467         GEM_BUG_ON(!vma->fence_size);
 468
 469         fenceable = (vma->node.size >= vma->fence_size &&
 470                      IS_ALIGNED(vma->node.start, vma->fence_alignment));
 471
 472         mappable = vma->node.start + vma->fence_size <= i915_vm_to_ggtt(vma->vm)->mappable_end;
 473
 474         if (mappable && fenceable)
 475                 vma->flags |= I915_VMA_CAN_FENCE;
 476         else
 477                 vma->flags &= ~I915_VMA_CAN_FENCE;
 478 }
 479
 480 static bool color_differs(struct drm_mm_node *node, unsigned long color)
 481 {
 482         return node->allocated && node->color != color;
 483 }
 484
 485 bool i915_gem_valid_gtt_space(struct i915_vma *vma, unsigned long cache_level)
 486 {
 487         struct drm_mm_node *node = &vma->node;
 488         struct drm_mm_node *other;
 489
 490         /*
 491          * On some machines we have to be careful when putting differing types
 492          * of snoopable memory together to avoid the prefetcher crossing memory
 493          * domains and dying. During vm initialisation, we decide whether or not
 494          * these constraints apply and set the drm_mm.color_adjust
 495          * appropriately.
 496          */
 497         if (vma->vm->mm.color_adjust == NULL)
 498                 return true;
 499
 500         /* Only valid to be called on an already inserted vma */
 501         GEM_BUG_ON(!drm_mm_node_allocated(node));
 502         GEM_BUG_ON(list_empty(&node->node_list));
 503
 504         other = list_prev_entry(node, node_list);
 505         if (color_differs(other, cache_level) && !drm_mm_hole_follows(other))
 506                 return false;
 507
 508         other = list_next_entry(node, node_list);
 509         if (color_differs(other, cache_level) && !drm_mm_hole_follows(node))
 510                 return false;
 511
 512         return true;
 513 }
 514
 515 static void assert_bind_count(const struct drm_i915_gem_object *obj)
 516 {
 517         /*
 518          * Combine the assertion that the object is bound and that we have
 519          * pinned its pages. But we should never have bound the object
 520          * more than we have pinned its pages. (For complete accuracy, we
 521          * assume that no else is pinning the pages, but as a rough assertion
 522          * that we will not run into problems later, this will do!)
 523          */
 524         GEM_BUG_ON(atomic_read(&obj->mm.pages_pin_count) < atomic_read(&obj->bind_count));
 525 }
 526
 527 /**
 528  * i915_vma_insert - finds a slot for the vma in its address space
 529  * @vma: the vma
 530  * @size: requested size in bytes (can be larger than the VMA)
 531  * @alignment: required alignment
 532  * @flags: mask of PIN_* flags to use
 533  *
 534  * First we try to allocate some free space that meets the requirements for
 535  * the VMA. Failiing that, if the flags permit, it will evict an old VMA,
 536  * preferrably the oldest idle entry to make room for the new VMA.
 537  *
 538  * Returns:
 539  * 0 on success, negative error code otherwise.
 540  */
 541 static int
 542 i915_vma_insert(struct i915_vma *vma, u64 size, u64 alignment, u64 flags)
 543 {
 544         struct drm_i915_private *dev_priv = vma->vm->i915;
 545         unsigned int cache_level;
 546         u64 start, end;
 547         int ret;
 548
 549         GEM_BUG_ON(i915_vma_is_closed(vma));
 550         GEM_BUG_ON(vma->flags & (I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
 551         GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
 552
 553         size = max(size, vma->size);
 554         alignment = max(alignment, vma->display_alignment);
 555         if (flags & PIN_MAPPABLE) {
 556                 size = max_t(typeof(size), size, vma->fence_size);
 557                 alignment = max_t(typeof(alignment),
 558                                   alignment, vma->fence_alignment);
 559         }
 560
 561         GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE));
 562         GEM_BUG_ON(!IS_ALIGNED(alignment, I915_GTT_MIN_ALIGNMENT));
 563         GEM_BUG_ON(!is_power_of_2(alignment));
 564
 565         start = flags & PIN_OFFSET_BIAS ? flags & PIN_OFFSET_MASK : 0;
 566         GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE));
 567
 568         end = vma->vm->total;
 569         if (flags & PIN_MAPPABLE)
 570                 end = min_t(u64, end, dev_priv->ggtt.mappable_end);
 571         if (flags & PIN_ZONE_4G)
 572                 end = min_t(u64, end, (1ULL << 32) - I915_GTT_PAGE_SIZE);
 573         GEM_BUG_ON(!IS_ALIGNED(end, I915_GTT_PAGE_SIZE));
 574
 575         /* If binding the object/GGTT view requires more space than the entire
 576          * aperture has, reject it early before evicting everything in a vain
 577          * attempt to find space.
 578          */
 579         if (size > end) {
 580                 DRM_DEBUG("Attempting to bind an object larger than the aperture: request=%llu > %s aperture=%llu\n",
 581                           size, flags & PIN_MAPPABLE ? "mappable" : "total",
 582                           end);
 583                 return -ENOSPC;
 584         }
 585
 586         if (vma->obj) {
 587                 ret = i915_gem_object_pin_pages(vma->obj);
 588                 if (ret)
 589                         return ret;
 590
 591                 cache_level = vma->obj->cache_level;
 592         } else {
 593                 cache_level = 0;
 594         }
 595
 596         GEM_BUG_ON(vma->pages);
 597
 598         ret = vma->ops->set_pages(vma);
 599         if (ret)
 600                 goto err_unpin;
 601
 602         if (flags & PIN_OFFSET_FIXED) {
 603                 u64 offset = flags & PIN_OFFSET_MASK;
 604                 if (!IS_ALIGNED(offset, alignment) ||
 605                     range_overflows(offset, size, end)) {
 606                         ret = -EINVAL;
 607                         goto err_clear;
 608                 }
 609
 610                 ret = i915_gem_gtt_reserve(vma->vm, &vma->node,
 611                                            size, offset, cache_level,
 612                                            flags);
 613                 if (ret)
 614                         goto err_clear;
 615         } else {
 616                 /*
 617                  * We only support huge gtt pages through the 48b PPGTT,
 618                  * however we also don't want to force any alignment for
 619                  * objects which need to be tightly packed into the low 32bits.
 620                  *
 621                  * Note that we assume that GGTT are limited to 4GiB for the
 622                  * forseeable future. See also i915_ggtt_offset().
 623                  */
 624                 if (upper_32_bits(end - 1) &&
 625                     vma->page_sizes.sg > I915_GTT_PAGE_SIZE) {
 626                         /*
 627                          * We can't mix 64K and 4K PTEs in the same page-table
 628                          * (2M block), and so to avoid the ugliness and
 629                          * complexity of coloring we opt for just aligning 64K
 630                          * objects to 2M.
 631                          */
 632                         u64 page_alignment =
 633                                 rounddown_pow_of_two(vma->page_sizes.sg |
 634                                                      I915_GTT_PAGE_SIZE_2M);
 635
 636                         /*
 637                          * Check we don't expand for the limited Global GTT
 638                          * (mappable aperture is even more precious!). This
 639                          * also checks that we exclude the aliasing-ppgtt.
 640                          */
 641                         GEM_BUG_ON(i915_vma_is_ggtt(vma));
 642
 643                         alignment = max(alignment, page_alignment);
 644
 645                         if (vma->page_sizes.sg & I915_GTT_PAGE_SIZE_64K)
 646                                 size = round_up(size, I915_GTT_PAGE_SIZE_2M);
 647                 }
 648
 649                 ret = i915_gem_gtt_insert(vma->vm, &vma->node,
 650                                           size, alignment, cache_level,
 651                                           start, end, flags);
 652                 if (ret)
 653                         goto err_clear;
 654
 655                 GEM_BUG_ON(vma->node.start < start);
 656                 GEM_BUG_ON(vma->node.start + vma->node.size > end);
 657         }
 658         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 659         GEM_BUG_ON(!i915_gem_valid_gtt_space(vma, cache_level));
 660
 661         mutex_lock(&vma->vm->mutex);
 662         list_move_tail(&vma->vm_link, &vma->vm->bound_list);
 663         mutex_unlock(&vma->vm->mutex);
 664
 665         if (vma->obj) {
 666                 atomic_inc(&vma->obj->bind_count);
 667                 assert_bind_count(vma->obj);
 668         }
 669
 670         return 0;
 671
 672 err_clear:
 673         vma->ops->clear_pages(vma);
 674 err_unpin:
 675         if (vma->obj)
 676                 i915_gem_object_unpin_pages(vma->obj);
 677         return ret;
 678 }
 679
 680 static void
 681 i915_vma_remove(struct i915_vma *vma)
 682 {
 683         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 684         GEM_BUG_ON(vma->flags & (I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
 685
 686         vma->ops->clear_pages(vma);
 687
 688         mutex_lock(&vma->vm->mutex);
 689         drm_mm_remove_node(&vma->node);
 690         list_move_tail(&vma->vm_link, &vma->vm->unbound_list);
 691         mutex_unlock(&vma->vm->mutex);
 692
 693         /*
 694          * Since the unbound list is global, only move to that list if
 695          * no more VMAs exist.
 696          */
 697         if (vma->obj) {
 698                 struct drm_i915_gem_object *obj = vma->obj;
 699
 700                 atomic_dec(&obj->bind_count);
 701
 702                 /*
 703                  * And finally now the object is completely decoupled from this
 704                  * vma, we can drop its hold on the backing storage and allow
 705                  * it to be reaped by the shrinker.
 706                  */
 707                 i915_gem_object_unpin_pages(obj);
 708                 assert_bind_count(obj);
 709         }
 710 }
 711
 712 int __i915_vma_do_pin(struct i915_vma *vma,
 713                       u64 size, u64 alignment, u64 flags)
 714 {
 715         const unsigned int bound = vma->flags;
 716         int ret;
 717
 718         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 719         GEM_BUG_ON((flags & (PIN_GLOBAL | PIN_USER)) == 0);
 720         GEM_BUG_ON((flags & PIN_GLOBAL) && !i915_vma_is_ggtt(vma));
 721
 722         if (WARN_ON(bound & I915_VMA_PIN_OVERFLOW)) {
 723                 ret = -EBUSY;
 724                 goto err_unpin;
 725         }
 726
 727         if ((bound & I915_VMA_BIND_MASK) == 0) {
 728                 ret = i915_vma_insert(vma, size, alignment, flags);
 729                 if (ret)
 730                         goto err_unpin;
 731         }
 732         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 733
 734         ret = i915_vma_bind(vma, vma->obj ? vma->obj->cache_level : 0, flags);
 735         if (ret)
 736                 goto err_remove;
 737
 738         GEM_BUG_ON((vma->flags & I915_VMA_BIND_MASK) == 0);
 739
 740         if ((bound ^ vma->flags) & I915_VMA_GLOBAL_BIND)
 741                 __i915_vma_set_map_and_fenceable(vma);
 742
 743         GEM_BUG_ON(i915_vma_misplaced(vma, size, alignment, flags));
 744         return 0;
 745
 746 err_remove:
 747         if ((bound & I915_VMA_BIND_MASK) == 0) {
 748                 i915_vma_remove(vma);
 749                 GEM_BUG_ON(vma->pages);
 750                 GEM_BUG_ON(vma->flags & I915_VMA_BIND_MASK);
 751         }
 752 err_unpin:
 753         __i915_vma_unpin(vma);
 754         return ret;
 755 }
 756
 757 void i915_vma_close(struct i915_vma *vma)
 758 {
 759         struct drm_i915_private *i915 = vma->vm->i915;
 760         unsigned long flags;
 761
 762         GEM_BUG_ON(i915_vma_is_closed(vma));
 763
 764         /*
 765          * We defer actually closing, unbinding and destroying the VMA until
 766          * the next idle point, or if the object is freed in the meantime. By
 767          * postponing the unbind, we allow for it to be resurrected by the
 768          * client, avoiding the work required to rebind the VMA. This is
 769          * advantageous for DRI, where the client/server pass objects
 770          * between themselves, temporarily opening a local VMA to the
 771          * object, and then closing it again. The same object is then reused
 772          * on the next frame (or two, depending on the depth of the swap queue)
 773          * causing us to rebind the VMA once more. This ends up being a lot
 774          * of wasted work for the steady state.
 775          */
 776         spin_lock_irqsave(&i915->gt.closed_lock, flags);
 777         list_add(&vma->closed_link, &i915->gt.closed_vma);
 778         spin_unlock_irqrestore(&i915->gt.closed_lock, flags);
 779 }
 780
 781 static void __i915_vma_remove_closed(struct i915_vma *vma)
 782 {
 783         struct drm_i915_private *i915 = vma->vm->i915;
 784
 785         if (!i915_vma_is_closed(vma))
 786                 return;
 787
 788         spin_lock_irq(&i915->gt.closed_lock);
 789         list_del_init(&vma->closed_link);
 790         spin_unlock_irq(&i915->gt.closed_lock);
 791 }
 792
 793 void i915_vma_reopen(struct i915_vma *vma)
 794 {
 795         __i915_vma_remove_closed(vma);
 796 }
 797
 798 static void __i915_vma_destroy(struct i915_vma *vma)
 799 {
 800         GEM_BUG_ON(vma->node.allocated);
 801         GEM_BUG_ON(vma->fence);
 802
 803         mutex_lock(&vma->vm->mutex);
 804         list_del(&vma->vm_link);
 805         mutex_unlock(&vma->vm->mutex);
 806
 807         if (vma->obj) {
 808                 struct drm_i915_gem_object *obj = vma->obj;
 809
 810                 spin_lock(&obj->vma.lock);
 811                 list_del(&vma->obj_link);
 812                 rb_erase(&vma->obj_node, &vma->obj->vma.tree);
 813                 spin_unlock(&obj->vma.lock);
 814         }
 815
 816         i915_active_fini(&vma->active);
 817
 818         i915_vma_free(vma);
 819 }
 820
 821 void i915_vma_destroy(struct i915_vma *vma)
 822 {
 823         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 824
 825         GEM_BUG_ON(i915_vma_is_pinned(vma));
 826
 827         __i915_vma_remove_closed(vma);
 828
 829         WARN_ON(i915_vma_unbind(vma));
 830         GEM_BUG_ON(i915_vma_is_active(vma));
 831
 832         __i915_vma_destroy(vma);
 833 }
 834
 835 void i915_vma_parked(struct drm_i915_private *i915)
 836 {
 837         struct i915_vma *vma, *next;
 838
 839         spin_lock_irq(&i915->gt.closed_lock);
 840         list_for_each_entry_safe(vma, next, &i915->gt.closed_vma, closed_link) {
 841                 list_del_init(&vma->closed_link);
 842                 spin_unlock_irq(&i915->gt.closed_lock);
 843
 844                 i915_vma_destroy(vma);
 845
 846                 spin_lock_irq(&i915->gt.closed_lock);
 847         }
 848         spin_unlock_irq(&i915->gt.closed_lock);
 849 }
 850
 851 static void __i915_vma_iounmap(struct i915_vma *vma)
 852 {
 853         GEM_BUG_ON(i915_vma_is_pinned(vma));
 854
 855         if (vma->iomap == NULL)
 856                 return;
 857
 858         io_mapping_unmap(vma->iomap);
 859         vma->iomap = NULL;
 860 }
 861
 862 void i915_vma_revoke_mmap(struct i915_vma *vma)
 863 {
 864         struct drm_vma_offset_node *node = &vma->obj->base.vma_node;
 865         u64 vma_offset;
 866
 867         lockdep_assert_held(&vma->vm->mutex);
 868
 869         if (!i915_vma_has_userfault(vma))
 870                 return;
 871
 872         GEM_BUG_ON(!i915_vma_is_map_and_fenceable(vma));
 873         GEM_BUG_ON(!vma->obj->userfault_count);
 874
 875         vma_offset = vma->ggtt_view.partial.offset << PAGE_SHIFT;
 876         unmap_mapping_range(vma->vm->i915->drm.anon_inode->i_mapping,
 877                             drm_vma_node_offset_addr(node) + vma_offset,
 878                             vma->size,
 879                             1);
 880
 881         i915_vma_unset_userfault(vma);
 882         if (!--vma->obj->userfault_count)
 883                 list_del(&vma->obj->userfault_link);
 884 }
 885
 886 int i915_vma_move_to_active(struct i915_vma *vma,
 887                             struct i915_request *rq,
 888                             unsigned int flags)
 889 {
 890         struct drm_i915_gem_object *obj = vma->obj;
 891         int err;
 892
 893         assert_vma_held(vma);
 894         assert_object_held(obj);
 895         GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 896
 897         /*
 898          * Add a reference if we're newly entering the active list.
 899          * The order in which we add operations to the retirement queue is
 900          * vital here: mark_active adds to the start of the callback list,
 901          * such that subsequent callbacks are called first. Therefore we
 902          * add the active reference first and queue for it to be dropped
 903          * *last*.
 904          */
 905         err = i915_active_ref(&vma->active, rq->timeline, rq);
 906         if (unlikely(err))
 907                 return err;
 908
 909         if (flags & EXEC_OBJECT_WRITE) {
 910                 if (intel_frontbuffer_invalidate(obj->frontbuffer, ORIGIN_CS))
 911                         i915_active_ref(&obj->frontbuffer->write,
 912                                         rq->timeline,
 913                                         rq);
 914
 915                 dma_resv_add_excl_fence(vma->resv, &rq->fence);
 916                 obj->write_domain = I915_GEM_DOMAIN_RENDER;
 917                 obj->read_domains = 0;
 918         } else {
 919                 err = dma_resv_reserve_shared(vma->resv, 1);
 920                 if (unlikely(err))
 921                         return err;
 922
 923                 dma_resv_add_shared_fence(vma->resv, &rq->fence);
 924                 obj->write_domain = 0;
 925         }
 926         obj->read_domains |= I915_GEM_GPU_DOMAINS;
 927         obj->mm.dirty = true;
 928
 929         GEM_BUG_ON(!i915_vma_is_active(vma));
 930         return 0;
 931 }
 932
 933 int i915_vma_unbind(struct i915_vma *vma)
 934 {
 935         int ret;
 936
 937         lockdep_assert_held(&vma->vm->i915->drm.struct_mutex);
 938
 939         /*
 940          * First wait upon any activity as retiring the request may
 941          * have side-effects such as unpinning or even unbinding this vma.
 942          */
 943         might_sleep();
 944         if (i915_vma_is_active(vma)) {
 945                 /*
 946                  * When a closed VMA is retired, it is unbound - eek.
 947                  * In order to prevent it from being recursively closed,
 948                  * take a pin on the vma so that the second unbind is
 949                  * aborted.
 950                  *
 951                  * Even more scary is that the retire callback may free
 952                  * the object (last active vma). To prevent the explosion
 953                  * we defer the actual object free to a worker that can
 954                  * only proceed once it acquires the struct_mutex (which
 955                  * we currently hold, therefore it cannot free this object
 956                  * before we are finished).
 957                  */
 958                 __i915_vma_pin(vma);
 959                 ret = i915_active_wait(&vma->active);
 960                 __i915_vma_unpin(vma);
 961                 if (ret)
 962                         return ret;
 963         }
 964         GEM_BUG_ON(i915_vma_is_active(vma));
 965
 966         if (i915_vma_is_pinned(vma)) {
 967                 vma_print_allocator(vma, "is pinned");
 968                 return -EBUSY;
 969         }
 970
 971         if (!drm_mm_node_allocated(&vma->node))
 972                 return 0;
 973
 974         if (i915_vma_is_map_and_fenceable(vma)) {
 975                 /*
 976                  * Check that we have flushed all writes through the GGTT
 977                  * before the unbind, other due to non-strict nature of those
 978                  * indirect writes they may end up referencing the GGTT PTE
 979                  * after the unbind.
 980                  */
 981                 i915_vma_flush_writes(vma);
 982                 GEM_BUG_ON(i915_vma_has_ggtt_write(vma));
 983
 984                 /* release the fence reg _after_ flushing */
 985                 mutex_lock(&vma->vm->mutex);
 986                 ret = i915_vma_revoke_fence(vma);
 987                 mutex_unlock(&vma->vm->mutex);
 988                 if (ret)
 989                         return ret;
 990
 991                 /* Force a pagefault for domain tracking on next user access */
 992                 mutex_lock(&vma->vm->mutex);
 993                 i915_vma_revoke_mmap(vma);
 994                 mutex_unlock(&vma->vm->mutex);
 995
 996                 __i915_vma_iounmap(vma);
 997                 vma->flags &= ~I915_VMA_CAN_FENCE;
 998         }
 999         GEM_BUG_ON(vma->fence);
1000         GEM_BUG_ON(i915_vma_has_userfault(vma));
1001
1002         if (likely(!vma->vm->closed)) {
1003                 trace_i915_vma_unbind(vma);
1004                 vma->ops->unbind_vma(vma);
1005         }
1006         vma->flags &= ~(I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND);
1007
1008         i915_vma_remove(vma);
1009
1010         return 0;
1011 }
1012
1013 struct i915_vma *i915_vma_make_unshrinkable(struct i915_vma *vma)
1014 {
1015         i915_gem_object_make_unshrinkable(vma->obj);
1016         return vma;
1017 }
1018
1019 void i915_vma_make_shrinkable(struct i915_vma *vma)
1020 {
1021         i915_gem_object_make_shrinkable(vma->obj);
1022 }
1023
1024 void i915_vma_make_purgeable(struct i915_vma *vma)
1025 {
1026         i915_gem_object_make_purgeable(vma->obj);
1027 }
1028
1029 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1030 #include "selftests/i915_vma.c"
1031 #endif
1032
1033 static void i915_global_vma_shrink(void)
1034 {
1035         kmem_cache_shrink(global.slab_vmas);
1036 }
1037
1038 static void i915_global_vma_exit(void)
1039 {
1040         kmem_cache_destroy(global.slab_vmas);
1041 }
1042
1043 static struct i915_global_vma global = { {
1044         .shrink = i915_global_vma_shrink,
1045         .exit = i915_global_vma_exit,
1046 } };
1047
1048 int __init i915_global_vma_init(void)
1049 {
1050         global.slab_vmas = KMEM_CACHE(i915_vma, SLAB_HWCACHE_ALIGN);
1051         if (!global.slab_vmas)
1052                 return -ENOMEM;
1053
1054         i915_global_register(&global.base);
1055         return 0;
1056 }